Geometry-aware PointNet for rapid prediction of cerebral aneurysm hemodynamics

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computational Fluid Dynamics

Background:

• Cerebral aneurysms affect 2-5% of the global population, posing a significant rupture risk.
• Computational fluid dynamics (CFD) offers detailed hemodynamic insights but is computationally intensive.
• Current methods limit routine clinical application of CFD for aneurysm risk assessment.

Purpose of the Study:

• To develop a deep learning model for rapid prediction of 3D velocity fields and wall shear stress (WSS) at peak systole.
• To create a fast, geometry-aware surrogate model for hemodynamic analysis in cerebral aneurysms.
• To overcome the computational limitations of traditional CFD for clinical risk stratification.

Main Methods:

• A dataset of 984 idealized middle cerebral artery bifurcation aneurysms was synthesized.
• CFD simulations generated ground-truth peak-systolic hemodynamic data.
• A point-cloud network with a distance-to-wall feature was developed to predict velocity and WSS.

Main Results:

• The model achieved high accuracy on idealized geometries: NMAE of 4.05% for velocity and 2.59% for WSS.
• Inference times were rapid: ~1.6 seconds for velocity and ~0.3 seconds for WSS.
• Significant performance degradation was observed on out-of-distribution non-idealized geometries.

Conclusions:

• The geometry-aware deep learning model enables fast hemodynamic prediction for idealized aneurysm geometries.
• Broader patient-specific training data and physiological boundary conditions are necessary for clinical translation.
• Further reliability assessments are required before routine clinical implementation of this AI tool.